PSI - Issue 51

Kenichi Ishihara et al. / Procedia Structural Integrity 51 (2023) 62–68 K. Ishihara and T. Meshii / Structural Integrity Procedia 00 (2022) 000–000

67

6

Thus, next and finally, the result when the default solver adam was changed to lbfgs is shown in Fig. 5 (b) and as Case C in Table 1. Adam refers to a stochastic gradient-based optimizer, and lbfgs is an optimizer in the family of quasi-Newton methods (Python, 2022). As a result, R 2 increased to the range of 0.96 to 0.98. Looking at the comparison between the true value of the learning data and the prediction results, the variation from the 45-degree line is smaller than the results of Case B were obtained, and it is plotted on this line. As a result of the above, a combination of parameters of Case C were accepted at this work. 4. Validation of constructed ANN DL model The constructed ANN DL model was applied to arbitrary welding conditions, and the validity of the DL model was confirmed to determine whether the target molten shape was obtained. As input data, v , Q , W and D not included in the train data and test data were set to the values and applied them to the constructed DL model. The CF equation and a general-purpose FEA solver LS-DYNA were used to validate the DL model. Input data and parameters a , b predicted by the constructed DL model and the both results were obtained for the target molten shape are shown in Table 2, and the comparison of the temperature contour by CF equation and FEA are shown in Fig. 6. Above the results, it can be said that the target molten shape was mostly expressed. Since the molten shape was approximately reproduced by the constructed DL model, the residual stress distribution can be calculated with high accuracy by using this TDA results.

Table 2 Input data and the parameter predicted by the constructed DL model and the molten shape obtained by CF equation and FEA.

Results

Input data (Feature)

Predicted parameter

CF equation

FEA

v [mm/s] Q [W] W [mm] D [mm]

a [mm]

b [mm]

W [mm]

D [mm]

W [mm]

D [mm]

5.0

6000

5

5

7.2

7.7

4.71

4.86

4.70

4.84

Fig. 6. The comparison of the temperature contour by CF equation and FEA.

5. Conclusion In this work, inverse analysis method of the molten shape was proposed by deep learning. As a result, deep learning model (DLM) was constructed to directly obtain heat source parameters necessary for stress analysis during welding process. This DLM can reduce the time required for parameters identification in FEA based failure analysis for welded components.